Slash commands that drive Deep Steve.

Describe it, file it, start it

Describe the problem in a sentence. The agent drafts a title and body, files the issue with the GitHub CLI, returns the URL, and offers to start work on it.

Accept, and Deep Steve opens a session preloaded with the issue in a dedicated worktree, leaving your main checkout untouched.

Merge and close out

Merges your branch into whichever branch the main checkout currently has checked out, so a merge follows your feature branch instead of always targeting main. Name a branch to override.

Commits any outstanding work first, closes the originating GitHub issue when there is one, and closes the session once the merge lands.

Summarize the work, then merge or stop

Summarizes what the worktree accomplished, including commits since the target branch and anything still uncommitted, then asks whether to merge.

Answer yes and it hands off to /deepsteve:merge. Answer no and it stops cleanly, leaving the branch and your uncommitted work untouched.

Two agents, one history

Forks the conversation into a second tab. Both sides carry the full history up to that point and continue independently, which lets you try a risky approach without abandoning the one that already works.

A plain shell in the session's directory

Opens a normal terminal tab with no agent in it, in the current working directory, optionally running a command at startup. Intended for long-lived processes: a dev server, a file watcher, a log tail.

The tab stays open until you close it. To run a single command and capture its output, the agent uses an MCP tool that tears its own tab down afterward.

Put an agent in a channel

Joins a Deep Steve chat channel, reads the history, and replies when a message is directed at it. Pass context to shape how it answers.

Pair it with /loop to keep watch: /loop 5m /deepsteve:chat #support.

Show it, don't print it

For an answer that is better looked at than read: the agent writes a page — a chart, a live dashboard — and opens it as its own tab beside the terminals. Ask for a change and it rewrites that same tab.

One-shot by design: closing the tab destroys it. For a page the project keeps across sessions, the agent registers a project mod instead.

Optimize against a metric, unattended

For optimization and algorithm-design problems. Describe what you are trying to improve and the agent identifies the pieces: the file being iterated on, the harness that scores it, the metrics, the baseline, and the finishing condition.

It files the plan as an issue, then runs the loop: change, measure, keep what wins, and stop when the finishing condition is met.